Physics-based ML model aims to predict microstructures in 3D printing faster

A project funded by the U.S. National Science Foundation at Lehigh University aims to calculate predictions of microstructure formation in metal 3D printing much more efficiently. The research group led by Parisa Khodabakhshi, Assistant Professor of Mechanical Engineering and Mechanics, is developing a physics-based, data-driven reduced-order model (ROM) that captures the formation of microstructures during binary alloy solidification and thus links process parameters directly to desired material properties. The funding period is three years, with a volume of USD 350,000.

Additive manufacturing produces parts layer by layer and enables geometries that are often difficult to access conventionally. At the same time, a variety of parameters — such as laser power, scan strategy, and cooling rates — influence the thermomechanical properties of the final part.

“This layer-by-layer approach allows for the fabrication of parts with complex geometries that are often difficult, or even impossible, to achieve with conventional manufacturing methods,” says Parisa Khodabakhshi, an assistant professor of Mechanical Engineering and Mechanics at Lehigh University’s P.C. Rossin College of Engineering and Applied Science. “However, the thermomechanical properties of the final additively manufactured parts are influenced by a large number of process parameters, making design optimization particularly challenging.”

Establishing the map between variations in process parameters and the final part’s properties requires several simulations across a wide range of length scales, making the task computationally expensive. “The computational demands of performing all the necessary simulations make it impractical,” says Khodabakhshi. As a result, manufacturers often resort to trial-and-error methods to achieve desired thermal or mechanical properties in the end product. “However, you cannot fully explore the entire design space that way to find the optimal design, which is why we’re currently not able to utilize the full potential of additive manufacturing.”

The planned ROM is intended to deliver the so-called forward map — from process parameters to solidification microstructure and then to part properties — approximately but significantly faster. Building on this, an inverse map can be constructed that infers parameter combinations from required properties. Methodologically, the team relies on a scientific machine learning framework that explicitly couples learning algorithms to conservation laws and boundary conditions.

“For example, say I want a part that has specific thermal properties,” she says. “I don’t know what my process parameters should be to achieve those properties. The simulations that link given process parameters to the resulting solidification microstructure, and consequently the final properties of the built part, are highly nonlinear. We refer to this simulation as the forward map. From there, I can construct the inverse map, which connects desired properties back to the process parameters.” The NSF project focuses on developing a computationally efficient model for the process-structure (PS) relationship.

Her team’s approach uses a scientific machine learning framework that blends data-driven machine learning algorithms with physical laws. “As engineers, we don’t want to just train a black-box algorithm,” says Khodabakhshi. “We want to embed physics into the problem to satisfy the governing equations of the physical phenomena so that we’re confident about the output that we receive from the algorithm. That’s the difference between conventional machine learning and scientific machine learning.”

The goal is more robust process design for industries with high qualification requirements such as aerospace, automotive, and medical technology. If the reduction in computing cost succeeds, design studies could be conducted more often with simulation support and less empirically.